Panel - Moderator · Moderator Xinli Gu, Huawei Technologies, USA Panelists Jos van Rooyen,...
Transcript of Panel - Moderator · Moderator Xinli Gu, Huawei Technologies, USA Panelists Jos van Rooyen,...
HUAWEI TECHNOLOGIES CO., LTD.
www.huawei.com
VALID’2016
ModeratorXinli Gu, Huawei Technologies, USA
Panelists
Jos van Rooyen, Indentify Test Services, the Nederland
Martin Zinner, Technische Universitat Dresden, Germany
Yingzhen Qu, Huawei Technologies, USA
Simon Genser, Kompetenzzentrum das Virtuelle Fahrzeug, Austria
Volker Gollucke, OFFIS Institute for Information Technology, Germany
Panel: Systems & Data – Challenges in Simulation and
Verification/Validation for Complex Systems & Big-Data
7. December 2011
HUAWEI TECHNOLOGIES Co., Ltd.
www.huawei.com
HUAWEI Confidential
Security Level: confidential
Thank You
www.huawei.com
© VIRTUAL VEHICLE
VIRTUAL VEHICLE Research Center is funded within the COMET – Competence Centers for Excellent Technologies – programme by the Austrian Federal
Ministry for Transport, Innovation and Technology (BMVIT), the Federal Ministry of Science, Research and Economy (BMWFW), the Austrian Research
Promotion Agency (FFG), the province of Styria and the Styrian Business Promotion Agency (SFG). The COMET programme is administrated by FFG.
Derivative Estimation via DELC-Method with Noisy
Data
Simon Genser
VIRTUAL VEHICLE Research Center
© VIRTUAL VEHICLE
My Person
10.10.2017 Derivative Estimation by DELC with Noisy Data 2
• Simon Genser
• Completed Master in Industrial Mathematics at the TU Graz, Austria.
• numerical methods (FEM, BEM, FDM,…)
• functional analysis
• partial differential equations
• Employed at the Virtual Vehicle Competence Center in Graz, Austria.
• Junior Researcher
• working on my PhD thesis about coupling of control systems
• Fields of Interest:
• numerical derivative estimation
• coupling of control systems
© VIRTUAL VEHICLE
Example: Sensitivity Analysis of a Blackbox System
10.10.2017 Derivation of a New Method for Derivative Estimation by Linear Combinations 3
Interested in Output-Input Sensitivity:
So we need:
Computation:
© VIRTUAL VEHICLE
Derivative Estimation by Linear Combinations (DELC) Method
10.10.2017 Derivation of a New Method for Derivative Estimation by Linear Combinations 4
Computation formula:
First solve:
© VIRTUAL VEHICLE
General Statement of the Problem
10.10.2017 Derivative Estimation by DELC with Noisy Data 5
One is interested in only by samples .
Therefore use the DELC method
For noiseless signals it works
but for noisy signals it fails.
© VIRTUAL VEHICLE
DELC with noisy signal
10.10.2017 Derivative Estimation by DELC with Noisy Data 6
© VIRTUAL VEHICLE
Why is the DELC Method not applicable for noisy signals?
10.10.2017 Derivative Estimation by DELC with Noisy Data 7
1. You can see the DELC Method as interpolating the samples
with a polynomial and derive this polynomial to get the derivative .
2. Affine Arithmetic Theory:
Noisy samples are represented as
with as the amplitude of the white noise.
The uncertainty in the DELC computation is
© VIRTUAL VEHICLE
Affine Arithmetic and DELC with noisy signal
10.10.2017 Derivative Estimation by DELC with Noisy Data 8
© VIRTUAL VEHICLE
How to improve it?
10.10.2017 Derivative Estimation by DELC with Noisy Data 9
1. Smooth the signal first and then use DELC on the smoothed signal
1. Smooth with Least Square Methods
2. ….
2. Some adjustment to the DELC to get rid of the interpolating behaviour
1. Try something like approximation instead of interpolation
2. Other ideas=?
Security Level:
www.huawei.com
Intelligent Data Center
with Machine Learning
Yingzhen Qu
Senior Research Scientist, Future Networks
Huawei Technologies, US Research Center
October 2017
HUAWEI TECHNOLOGIES CO., LTD. Page 2
A Generalized Machine Learning Loop
Examples RegressionNeural Networks
outputinput PredictionA
pply
Reward(Data, action-A)Penalize(data, action-B)
input
HUAWEI TECHNOLOGIES CO., LTD. Page 3
An Example: ECMP Based Link Utilization Problem in a Switch
Massive Scale DCs use fixed spine-leaf topology
ECMP distributes traffic across multiple paths
ECMP uses Hash computation to balance similar flows over multiple links
However, the flows are not evenly balanced
› Low-bandwidth (Mice) flows: Majority of flows are short-lived and latency sensitive.
» Example: Web, chat applications
› High-bandwidth (Elephant) flows consume majority bandwidth and are long-lived.
» Example Storage-intensive big-data, data-replication and backup applications
Problem
› Variance in the amount of bandwidth used between long-lived vs short-lived flows does not ensure that traffic
is balanced across all the links.
› Increase in Mean-time-to completion for mice flows
› Reduced data-rate for elephant flows due to congestion control
HUAWEI TECHNOLOGIES CO., LTD. Page 4
Facebook Data Center Fabric
Source: https://code.facebook.com/posts/360346274145943/introducing-data-center-fabric-the-next-generation-facebook-data-center-network/
HUAWEI TECHNOLOGIES CO., LTD. Page 5
Machine Learning for ECMP Link Utilization in a Switch
ECMP linksElephant flows
Mice flowssame hash forcertain flows
outputinput PredictionMiceflow?
Y
ECMP Link select
No
• 70% times pickECMP prescribedlink
• 30% Pick LeastCongested Link
Live input
Classify flow Tuple actions
Classify flow size > 2Mbps as elephant flows
RLA – Reinforcement Learning Algorithm
HUAWEI TECHNOLOGIES CO., LTD. Page 6
Intelligence Driven Networking – DC Scenarios with Global Scope
• Extend to wider scoped learning - Global models acrossmultiple switches
• Different Learning models for different scenariostogether
Src: Facebook data center
Learn to optimally place services, VMs in DC
Optimize End to End paths
© Identify 2009-2017 13-11-2017
IT mediation
Jos van Rooyen
© Identify 2009-2017
• Employed at Identify as partner / principal consultant
• >25 years in software testing & quality management
• Co-author TestGrip ,TestFrame, Project de Baas, QualitySupervision, Textbook; “Aan de slag met software testen”
• Test expert online magazine Computable
• Publication areas; Test process Improvement, BI-testing, Testautomation, Test Education, Risk Based Testing, QualitySupervision
• Review committee software conference Valid2017
• Founder of the “Houten Groep”
• Member NESMA working party; Metrics in Contracts
• Member of several working parties Dutch Testing Society:
• Model Based Testing
• Test Education universities of applied science
22
Who am I?
© Identify 2009-2017
Contracts by suppliers are very promising
A lot of functionality for a low price
I see a lot of projects failing just at this point:
• Wrong functionality or less functionality delivered
• Misinterpretation of the requirements
• Delivery of the system too late
• Budget problems. Running out of money
• Poor quality so the demanding organization has todo a lot of rework
• Demand / supply doesn’t work together
33
Interpretation of the problem
© Identify 2009-2017
Based on IT mediation:
Determination of the suitability of the new information system in theexisting organization, with respect to the business process, the people, ITand the organization:
• Quality
• Maintainability
• Metrics
• Process
• Vendor management
• Contract management
• Legal
44
Possible solutions to solve the problem
© Identify 2009-2017
• Based on metrics:
• Number of defects discovered
• Defect Removal Efficiency
• Compliance to the requested functional quality
• Quality in use
• Coverage degree requirements
55
Possible solutions to solve the problem(2)
© Identify 2009-2017
• Do you recognize the described situation?
• What are your ideas to manage these kind of situations?
• Are the described solutions indeed good solutions?
66
Questions to the audience
© Identify 2009-2017 Ketenregie
Jos van RooyenE-mail: [email protected]: +31-654906282LinkedIn: Jos van Rooyen
Questions?
Data Analysis for a Reference Waterway
Dr.-Ing. Volker Gollücke
Observe, , Test
Reference Waterway
The Reference Waterway covers the Elbeand Kiel Canal Approach, Germany. Itcovers a basic maritime surveillanceinfrastructure with three Sensorboxes(including AIS, Radar, cameras) and broadband communication.
• Setting up a databasewith travel pattern andnear collisions
• Gathering Live Radar,AIS, Wind and VideoData
• Connected to the eMirInfrastructure with LTE
Observing and learning from the real world
13.11.2017
All sensors can be gathered foran overview of the situation onthe ElbeeMir Infrastructure allows dataexchange between differentsources and users
Example: Maneuverpoints
13.11.2017
application of the knowledge acquired
13.11.2017
• Research Vessel• Not only AIS data but also Video Streams• Online prediction of behavior of target vessels
Challenges Little’s Law Operating Curve Management Conclusion Literature
Manufacturing AnalyticsMethods to analyze and improve the performance of a fab
Martin Zinner
October, 10 2017 @ ICSEA 2017 (Athens)
1 / 14
Challenges Little’s Law Operating Curve Management Conclusion Literature
Agenda
Challenges
Little’s Law
Operating Curve Management
Conclusion
Literature
2 / 14
Challenges Little’s Law Operating Curve Management Conclusion Literature
Agenda
Challenges
Little’s Law
Operating Curve Management
Conclusion
Literature
3 / 14
Challenges Little’s Law Operating Curve Management Conclusion Literature
Challenges in a fab
I Expensive tools should not lack of material to process.I Tools must be maintained having downtimes (planned and
unplanned).
Technical Objectives:1) Reduce inventory and implicitly reduce cycle time.2) Increase throughput.
4 / 14
Challenges Little’s Law Operating Curve Management Conclusion Literature
Agenda
Challenges
Little’s Law
Operating Curve Management
Conclusion
Literature
5 / 14
Challenges Little’s Law Operating Curve Management Conclusion Literature
Little’s Law
I Average cycle time (“flow time” / “sojourn time”)(avg)CT : Time interval that a unit spent in the system
I Average work in process - inventory(avg)WIP: Average number of items in the system
I ThroughputTH: Average number of items arriving per unit time
The average departure rate exists and equals the average arrival rate:
(avg)CT =(avg)WIP
TH.
6 / 14
Challenges Little’s Law Operating Curve Management Conclusion Literature
Little’s Law: Valability Conditions
Valability Conditions:I In steady state and non preemptive systems (properties are
unchanged in time, no interrupts and later resumes)I In stochastic systems if the probabilities that some events
occur in the system are constantI Valid as limit on time (holds eventually)
Properties:I Average departure rate exists, equals the average arrival rate.I It can be used on all level of the production system (tools,
single production step, group of production steps, wholeproduction line for a product, whole production line).
I It is one of the central theorems of the operational research.
7 / 14
Challenges Little’s Law Operating Curve Management Conclusion Literature
Agenda
Challenges
Little’s Law
Operating Curve Management
Conclusion
Literature
8 / 14
Challenges Little’s Law Operating Curve Management Conclusion Literature
Flow Factor
Flow factor FF is a reference value, controls the production.
I Cycle time CT : Time to process unit including waiting timeI Raw process time RPT : Minimal time to process a unit
FF :=CT
RPT.
FF can be calculated by . . .a) . . . applying Little’s Law to estimate the cycle time,b) . . . using the effective values of the cycle time.
9 / 14
Challenges Little’s Law Operating Curve Management Conclusion Literature
Capacity, Utilization, Variability [3]
I Capacity Capa: Maximal throughputI Utilization U: Throughput / CapacityI Variability α: Can be calculated using a single (FF ,U) tuple
FF = α · UU − 1
+ 1.
Utilization vs. flow factor Throughput vs. cycle time10 / 14
Challenges Little’s Law Operating Curve Management Conclusion Literature
Agenda
Challenges
Little’s Law
Operating Curve Management
Conclusion
Literature
11 / 14
Challenges Little’s Law Operating Curve Management Conclusion Literature
Conclusion
Requested optimization objectives:1) Reduce inventory and implicitly reduce the cycle time.2) Increase throughput.
Conclusion:The two objectives are contradictory!They cannot be optimized at the same time!
Challenge:Seeking a good balance between both objectives.
12 / 14
Challenges Little’s Law Operating Curve Management Conclusion Literature
Agenda
Challenges
Little’s Law
Operating Curve Management
Conclusion
Literature
13 / 14
Challenges Little’s Law Operating Curve Management Conclusion Literature
Literature
1) Book, one of the best:Wallace J. Hopp, Marc L. Spearman: “Factory Physics:Foundations of Manufacturing Management”, 2nd edition,2001, McGraw-Hill, Higher Education
2) PhD thesis, easy to follow:Kirsten Hilsenbeck: “Optimierungsmodelle in derHalbleiterproduktionstechnik”, 2005, Doctoral dissertation,Technische Universität München,https://mediatum.ub.tum.de/doc/601620/601620.pdf
3) Slides, very accurate and comprehensive:Walter Hansch, Tina Kubot: ‘Factory Dynamics”, Chapter 7,Lectures at Universität der Bunderwehr Munich,http://www.unibw.de/rz/dokumente/fakultaeten/
getFILE?fid=6186607&tid=fakultaeten
14 / 14
Introduction Practical Example Summary Bibliography
Real-Time Data Analytics
Martin Zinner
October, 10 2017 @ ICSEA 2017 (Athens)
1 / 11
Introduction Practical Example Summary Bibliography
Agenda
Introduction
Practical Example
Summary
Bibliography
2 / 11
Introduction Practical Example Summary Bibliography
Agenda
Introduction
Practical Example
Summary
Bibliography
3 / 11
Introduction Practical Example Summary Bibliography
Research project “Cool Silicon”
1 out of 5 leading-edge clusterswithin hightech strategy of the German Federal Government
I Project term:2009 – 2014
I Budget:135 million Euro
Members of Cool Silicon (Source: cool-silicon.de)
Aim: Substantially increase energy efficiency of ICs andinformation technology
4 / 11
Introduction Practical Example Summary Bibliography
Scope of a IT-subproject in “Cool Silicon”
Aims:1) Performance improvement2) Reducing energy consumption3) New functionality for Business Intelligence
Methods:I Investigation and testing of appropriate infrastructure and
system architectureI Definition and testing of appropriate modeling / algorithmsI Creation of prerequisites for efficient real-time aggregations
and analysis (including ad hoc reporting)I Creating prerequisites for efficient parallelizationI Reduction of complexity of nightly batch runs for data
aggregation5 / 11
Introduction Practical Example Summary Bibliography
Agenda
Introduction
Practical Example
Summary
Bibliography
6 / 11
Introduction Practical Example Summary Bibliography
Illustration of principles of real-time data analytics
Practical example: Computation of standard deviation sN for Nobservations / items x1, x2, x3, . . . , xN
I Classical calculation: (Cannot be used for real-timecomputation)
sN =( 1
N
N∑i=1
(xi − x̄)2)1/2
with x̄ =1N
N∑i=1
xi .
To calculate sN red sum has to be recomputed from scratch.
I Real-time calculation:
sN =1N
(∣∣N N∑i=1
x2i − (
N∑i=1
xi)2∣∣)1/2
.
Obtained by multiplication and regrouping.If new item is added, then both sums can be updated.sN will be calculated only on demand. 7 / 11
Introduction Practical Example Summary Bibliography
Agenda
Introduction
Practical Example
Summary
Bibliography
8 / 11
Introduction Practical Example Summary Bibliography
Classical vs. real-time aggregation
Classical aggregation (batch mode):Calculation is started only when all the data involved becomesavailable (e.g., end of the day, etc.).
I Large amounts of data have to be calculated.I Potentially complicated performance improvement algorithms
are necessary.I Race against time.
Real-time aggregation:Data are aggregated as soon as data are known to the system.
I Small amounts of data are computed quasi continuously (e.g.,in memory spread over the whole day).
I Smaller hardware is sufficient.I Energy efficient.
9 / 11
Introduction Practical Example Summary Bibliography
Agenda
Introduction
Practical Example
Summary
Bibliography
10 / 11
Introduction Practical Example Summary Bibliography
Bibliography
Patent US 2017/0032016 A1Zinner et al.: “Real-time Information Systems and MethodologyBased on Continuous Homomorphic Processing in LinearInformation Spaces”http://www.google.ch/patents/US20170032016
11 / 11